Multiple files were hardcoding ~/.config/opencode paths instead of using
getOpenCodeConfigDir() which respects the OPENCODE_CONFIG_DIR env var.
This broke profile isolation features like OCX ghost mode, where users
set OPENCODE_CONFIG_DIR to a custom path but oh-my-opencode.json and
other configs weren't being read from that location.
Changes:
- plugin-config.ts: Use getOpenCodeConfigDir() directly
- cli/doctor/checks: Use getOpenCodeConfigDir() for auth and config checks
- tools/lsp/config.ts: Use getOpenCodeConfigDir() for LSP config paths
- command loaders: Use getOpenCodeConfigDir() for global command dirs
- hooks: Use getOpenCodeConfigDir() for hook config paths
- config-path.ts: Mark getUserConfigDir() as deprecated
- tests: Ensure OPENCODE_CONFIG_DIR is properly isolated in tests
OpenCode v1.1.30+ added HTTP-based SessionList/Read/Search/Info tools that
require localhost:4096 server. Renamed our tools to PascalCase to override
them with local file-based implementation, fixing 'Unable to connect' errors.
- Change skill directory from 'skill' (singular) to 'skills' (plural) to match OpenCode standard
- Use getOpenCodeConfigDir() to respect OPENCODE_CONFIG_DIR environment variable
- Update documentation and tests to reflect correct paths
Fixes#810
Previously, unstable agents (gemini models, is_unstable_agent=true) were
forced to background mode but returned immediately with task ID. This
caused callers to lose visibility into results.
Now: launch as background for monitoring stability, but poll and wait
for completion, returning actual task output like sync mode.
🤖 Generated with assistance of [OhMyOpenCode](https://github.com/code-yeongyu/oh-my-opencode)
Detects unstable agents via is_unstable_agent config or gemini in model
name. Forces background mode even when run_in_background=false, with
system message explaining the forced conversion for monitoring.
Previously, when using categories like 'quick', the parent session's model
(e.g., Opus 4.5) would override the category's built-in model (e.g., Haiku).
Fixed priority: userConfig.model → category built-in → systemDefault
The inherited model from parent session no longer affects category-based
delegation - categories have their own explicit models.
- Remove temperature from all categories
- Consolidate CATEGORY_MODEL_CATALOG into DEFAULT_CATEGORIES
- Replace 'general' and 'most-capable' with 'unspecified-low' and 'unspecified-high'
- Add Selection_Gate to unspecified categories to force deliberate selection
- Update quick category to use claude-haiku-4-5
- Update all references and tests across codebase
- Fixed delegate-task/tools.ts to access openCodeConfig.data.model instead of openCodeConfig.model
- Updated error messages to use getOpenCodeConfigPaths() for cross-platform paths
- Fixed 12 test mocks to use correct { data: { model: ... } } structure
- Fixes delegation failing with 'requires a default model' even when model is configured
BREAKING CHANGE: Model resolution overhauled
- Created centralized model-resolver.ts with priority chain:
userModel → inheritedModel → systemDefaultModel
- Removed model field from all 7 DEFAULT_CATEGORIES entries
- Removed DEFAULT_MODEL constants from 10 agents
- Removed singleton agent exports (use factories instead)
- Made CategoryConfigSchema.model optional
- CLI no longer generates model overrides
- Empty strings treated as unset (uses fallback)
Users must now:
1. Use factory functions (createOracleAgent, etc.) instead of singletons
2. Provide model explicitly or use systemDefaultModel
3. Configure category models explicitly if needed
Fixes model fallback bug where hardcoded defaults overrode
user's OpenCode configured model.
- Change skills type from string[] | null to string[]
- Allow skills=[] for no skills, reject skills=null
- Remove emojis from error messages and prompts
- Update tests accordingly
- Root AGENTS.md: Updated timestamp, commit hash, line counts
- src/agents/AGENTS.md: Updated to 50 lines, current structure
- src/cli/AGENTS.md: Updated to 57 lines, current structure
- src/features/AGENTS.md: Updated to 65 lines, current structure
- src/hooks/AGENTS.md: Updated to 53 lines, current structure
- src/shared/AGENTS.md: Updated to 52 lines, core utilities
- src/tools/AGENTS.md: Updated to 50 lines, tool categories
🤖 Generated with assistance of [OhMyOpenCode](https://github.com/code-yeongyu/oh-my-opencode)
- Changed lsp_diagnostics error handling to throw errors instead of returning strings
- Line 211: Changed from `return output` to `throw new Error(output)`
- Makes errors display as proper error blocks in TUI
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
Previously, parent model string would override category default model,
causing categories like 'ultrabrain' to use the parent's model (e.g., sonnet)
instead of the intended category default (e.g., gpt-5.2).
Model priority is now:
1. userConfig.model (oh-my-opencode.json override)
2. defaultConfig.model (category default)
3. parentModelString (fallback)
4. systemDefaultModel (last resort)
- Remove 'Start Sisyphus work session' text check, keep only <session-context> tag
- Update interactive_bash description with WARNING: TMUX ONLY emphasis
- Update tests to use <session-context> wrapper
OpenCode API returns different structures for user vs assistant messages:
- User: info.model = { providerID, modelID } (nested)
- Assistant: info.modelID, info.providerID (flat top-level)
Previous code only checked nested format, causing model info loss when
continuation hooks fired after assistant messages.
Files modified:
- todo-continuation-enforcer.ts
- ralph-loop/index.ts
- sisyphus-task/tools.ts
- background-agent/manager.ts
Added test for assistant message model extraction.
- Change skills type from string[] to string[] | null
- Empty array [] now returns error with available skills list
- null is allowed for tasks that genuinely need no skills
- Updated tests to use skills: null instead of skills: []
- Forces explicit decision: either specify skills or justify with null
The watermark (commit footer and co-author) was inconsistently applied because:
1. The skill tool didn't receive gitMasterConfig
2. The approach was 'default ON, inject DISABLED override' which LLMs sometimes ignored
This refactors to 'inject only when enabled' approach:
- Remove hardcoded watermark section from base templates
- Dynamically inject section 5.5 based on config values
- Default is still ON (both true when no config)
- When both disabled, no injection occurs (clean prompt)
Also fixes missing config propagation to skill tool and createBuiltinAgents.
Previously, continuation hooks (todo-continuation, boulder-continuation, ralph-loop)
and background tasks resolved model info from filesystem cache, which could be stale
or missing. This caused session.prompt to fallback to default model (Sonnet) instead
of using the originally configured model (e.g., Opus).
Now all session.prompt calls first try API (session.messages) to get current model
info, with filesystem as fallback if API fails.
Affected files:
- todo-continuation-enforcer.ts
- sisyphus-orchestrator/index.ts
- ralph-loop/index.ts
- background-agent/manager.ts
- sisyphus-task/tools.ts
- hook-message-injector/index.ts (export ToolPermission type)
Add formatDetailedError helper that includes:
- Full args dump (description, category, agent, skills)
- Session ID and agent info
- Stack trace (first 10 lines)
Applied to all catch blocks for better debugging.
- Update generated timestamp and commit hash metadata
- Normalize agent model names with provider prefixes (anthropic/, opencode/, google/)
- Remove deprecated Google OAuth/Antigravity references
- Update line counts and complexity hotspot entries
- Adjust test count from 82 to 80+ files and assertions from 2559+ to 2500+
🤖 Generated with assistance of oh-my-opencode
LLMs often call look_at with 'path' instead of 'file_path' parameter,
causing TypeError and infinite retry loops.
- Add normalizeArgs() to accept both 'path' and 'file_path'
- Add validateArgs() with clear error messages showing correct usage
- Add tests for normalization and validation
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
Rename to signal mutation behavior - the function advances the cursor
as a side effect, so 'consume' better reflects that calling it twice
with the same input yields different results.
Nguyen Khac Trung Kien